Speculative decoding repo compares proposer strategies

Hot take: this is most valuable as a systems explainer, not just a speed benchmark. It shows that speculative decoding performance is determined by the whole proposer-verifier pipeline, not by acceptance rate alone.

• –The shared contract makes it easier to compare proposer quality, verifier cost, and acceptance behavior across very different methods.
• –Learned heads like EAGLE-3 and Medusa-1 are fundamentally different from draft-model speculation because they attach speculation to internal model representations instead of relying on a separate smaller LM.
• –PARD is a good reminder that lower acceptance can still win if proposer generation is cheap enough and parallelism is better utilized.
• –N-gram and suffix decoding are useful baselines because they expose how far training-free reuse can go when prompts have repeated structure.
• –The repo’s small-slice benchmarks are best treated as implementation evidence and workflow intuition, not broad throughput claims.